Reducing image artifacts in sparse projection CT using conditional generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Reducing projection data in computed tomography (CT) lowers radiation dose but introduces image artifacts.
• Sparse-view CT (SVCT) presents challenges in maintaining diagnostic image quality due to artifacts.

Purpose of the Study:

• To quantitatively evaluate the effectiveness of conditional generative adversarial networks (CGAN) for restoring image quality in sparse-view CT.
• To compare the performance of CGAN against autoencoder (AE) and U-Net models in sparse-view CT image restoration.

Main Methods:

• Simulated sparse projection CT images were used to train AE, U-Net, and CGAN models.
• Models were trained on artifact-original image pairs, with 90% of patient cases for training and 10% for evaluation.
• Image quality was assessed using mean error (ME), mean absolute error (MAE), structural image similarity (SSIM), and peak signal-to-noise ratio (PSNR).

Main Results:

• All models showed image quality improvement, but AE and U-Net resulted in decreased resolution and blurring, with significant errors in lung/air regions.
• CGAN demonstrated accurate CT value restoration and superior SSIM and PSNR compared to AE and U-Net.
• Minor deformations in tumor/spine regions and some hallucination artifacts were noted in CGAN results.

Conclusions:

• CGAN models offer a promising approach for high-fidelity image quality restoration in sparse-view CT.
• CGAN outperforms AE and U-Net in mitigating artifacts and preserving image fidelity for low-dose CT applications.
• Further research may address minor artifacts observed in CGAN for complete artifact suppression in SVCT.